Prednosti in slabosti globokih nevronskih omrežij za razpoznavanje obrazov

BRUMEN, LUKA (author), Štruc, Vitomir (mentor)

V delu je analiziran vpliv spremembe osvetlitve, poze subjekta, velikost izreza obraznega področja, kvalitete in resolucije na razpoznavanje obrazov enega najmodernejših modelov globoke nevronske mreže. Rdeča nit dela je sposobnost verifikacije 16 plastnega modela konvolucijske nevronske mreže VGG, da določi ali je na paru slik ista oseba (uporabnik) ali ne (vsiljivec). Za potrebe nekaterih testov smo razvili postopke za simulacijo nenadzorovanih pogojev zajema slik in analizirali rezultate. Slike, ki smo jih pridobili z glajenjem, izrezovanjem obraznega področja in zniževanjem resolucije, lahko razumemo kot posledico slabših pogojev pri zajemu ali slabšega sistema za zajem slik, uporaba le teh v aplikacijah pa poceni končno ceno. Pokazali bomo, da se model v takšnih pogojih dobro obnese. Robustnost modela na različne svetlobne pogoje bomo preizkusili na zbirki EYB, medtem ko bomo robustnost modela na vpliv poze subjekta analizirali z zbirko FERET. Za ostale teste smo uporabili zbirko LFW, ki smo jo za potrebe tega dela obdelali tako, da smo postopoma slabšali slike uporabnikov. Razdalje med vektorji značilk smo računali s pomočjo kosinusne razdalje, za prikaz učinkovitosti smo uporabili krivulje ROC.

• Language: Slovenian
• Keywords: računalniški vid, razpoznavanje obrazov, globoke nevronske mreže, konvolucijske nevronske mreže, CNN, ConvNet
• Work type: Undergraduate thesis
• Organization: FE - Faculty of Electrical Engineering
• Year: 2016
• PID: 20.500.12556/RUL-85009
• Publication date in RUL: 09.09.2016

Secondary language

• Language: English
• Title: Strengths and weaknesses of deep face recognition models

Abstract

In this paper we analyse the effects of changes in illumination, pose, image size, facial area size, image quality and resolution on the face recognition performance of a state-of-the-art deep neural network. The main focus of this work lies in the performance of the VGG model, a 16-layer convolutional neural network and its ability to correctly classify several pairs of test images as clients or imposters. For some of the tests, we provide methods that simulate image acquisition in unstable and uncontrolled environments and discuss the results. Effects of blurring, cropping and resizing facial images could be understood as a result of a less powerful image acquisition system in terms of quality, the use of which consequently reduces the costs of a day to day face recognition application. We show that the said system performs well in such conditions. The effects of illumination and pose on face recognition accuracy are analysed using EYB and FERET datasets respectively. The study of other said changes of environmental variables are analysed on the LFW database and for which we pre-prepare several subsets, where we steadily deteriorate the conditions of client's probe images compared to the gallery image and measure the distance between its feature vectors using the cosine distance.

• Keywords: computer vision, face recognition, deep neural network, convolutional neural network, CNN, ConvNet